US 12,445,851 B2
Intrusion detection method and device for in-vehicle controller area network
Yong Qi, Jiangsu (CN); and Yangwei Sun, Jiangsu (CN)
Assigned to NANJING UNIVERSITY OF SCIENCE AND TECHNOLOGY, Nanjing (CN)
Appl. No. 18/577,181
Filed by NANJING UNIVERSITY OF SCIENCE AND TECHNOLOGY, Jiangsu (CN)
PCT Filed Feb. 23, 2023, PCT No. PCT/CN2023/077806
§ 371(c)(1), (2) Date Jan. 5, 2024,
PCT Pub. No. WO2023/160600, PCT Pub. Date Aug. 31, 2023.
Claims priority of application No. 202210165407.4 (CN), filed on Feb. 23, 2022.
Prior Publication US 2024/0224041 A1, Jul. 4, 2024
Int. Cl. H04W 12/121 (2021.01); G06N 20/00 (2019.01); H04L 12/40 (2006.01)
CPC H04W 12/121 (2021.01) [G06N 20/00 (2019.01); H04L 12/40 (2013.01); H04L 2012/40215 (2013.01)] 15 Claims
OG exemplary drawing
 
1. An intrusion detection method for an in-vehicle controller area network, comprising:
digitizing and normalizing collected original data, obtaining preprocessed data, and dividing the preprocessed data into a training set and a test set;
conducting feature selection on the preprocessed data through a particle swarm optimization (PSO)-light gradient boosting machine (GBM) bidirectional feature selection method; and
classifying test set data subjected to the feature selection with a stacking integrated model, and obtaining an intrusion detection result, wherein
the PSO-LightGBM bidirectional feature selection method comprises:
firstly conducting parameter optimization on a LightGBM with a PSO algorithm; then arranging feature importance in descending order with the LightGBM, selecting all sorted feature sets, deleting a feature having least importance from a current feature set each time such that a new feature subset is formed, conducting feature deletion on data according to the new feature subset, and conducting classification prediction by means of the stacking integrated model; cyclically deleting, if precision of a prediction result is not reduced, a feature having least importance, and conducting feature deletion on the new feature subset; and withdrawing, if precision of a prediction result is reduced, feature deletion, ending the feature deletion, and returning a data set containing only features after feature deletion.